Four Network Structures Could Explain Why Digital Systems Fail

A week of major outages has made digital infrastructure look more fragile than many people expected. Microsoft 365, banking services, air traffic systems, an airline, and the National Security Agency’s classified communications network all went down within days of one another.
That timing pushed conspiracy theories across social media, with television and YouTube commentators blaming foreign governments or rogue AI agents. But officials and researchers have pointed to routine maintenance, software defects, overloaded equipment, and weak connections between old and new systems. A new result in network theory offers another way to understand why failures can spread: four basic network structures can block a system from achieving “perfect resilience.”
What perfect resilience means
When links in a communication network fail, messages can get trapped in a loop instead of reaching their destinations. For more than a decade, computer scientists have tried to determine which networks can avoid that problem under realistic routing rules.
A paper presented at the Symposium on Theory of Computing in June 2026 settled the question with a comprehensive proof. It showed that exactly four minimal network structures can thwart perfect resilience. In networks that do not contain those structures, engineers can always create rules that reliably deliver packets from any source to any destination, no matter which links fail.
The result builds on work that began in 2012, when Schapira and five other researchers formally defined perfect resilience and created a mathematical model for studying when it is possible. Their model assumes that routers must make decisions with limited information. A router can choose where to send each new packet using only three factors: the packet’s destination, the link it arrived on, and which neighboring links have failed.
The rules cannot depend on the route a packet has already taken because routers do not have access to that history. In 2020, Stefan Schmid and his colleagues proved a partial result about network reliability. Christian Kielmann Godfrey, Schapira, and their colleagues later showed that perfect resilience is not always possible in this model by finding a single network layout that cannot be made perfectly resilient.
Michael Schapira, a computer scientist at the Hebrew University of Jerusalem who was not involved in the new result, called it “a very elegant contribution.” He added, “It’s pretty rare that you get a complete understanding of something.” Brighten Godfrey of the University of Illinois, Urbana-Champaign described the model as “something very close to how networks actually operate.”
A crowded week of failures
The real-world outages began before the theory drew attention. On October 28, a train crash in Spain was linked to a safety system that went down. Transportation Secretary Sean Duffy initially said outdated communications equipment had been overloaded with data. He later described a new router sending a flood of connection attempts every few seconds and swamping other routers.
The router was being installed by L3Harris, which was replacing copper wire with fiber at a facility in southern Arkansas. Robert N. Charette said the company was replacing old Cisco routers with new Juniper routers, and that the old Cisco equipment had been in place for about 20 years. He said the crashing system was actually taking down the old network, not the new routers.
Charette said the Arkansas location and another router in Amarillo, Texas, began pinging back and forth. He also said the Amarillo router had been installed three weeks earlier without a problem, while 400 routers had been installed so far without a problem and 69 remained operational. His conclusion was blunt: “That somebody didn’t do their homework.”
He connected the failure to configuration management and gaps in the operators’ understanding of the network. The FAA had also expanded its AI-based SMART traffic tool to some facilities, but the FAA and the tool’s developer denied that it caused the outage.
On Thursday, November 12th, Microsoft 365 Exchange Online, Outlook, Teams, and Copilot stopped working for roughly a day. Downdetector showed at least 74 countries affected. Microsoft said the incident was not a cyberattack and was not caused by AI agents. Instead, a latent defect in the networking gateway combined with a previous capacity upgrade and a recent network change, leaving servers unresponsive under certain load conditions.
The next day, the FAA ordered a full nationwide groundstop. Every airport in the United States stopped sending planes into the sky, while aircraft already in the air had to return to the ground. Friday’s groundstop was only the second since 9/11. On Monday, Chase Bank customers struggled to access their accounts for the better part of the day, United Airlines suffered an outage, and on Tuesday the NSA’s classified communications network went down.
Why the failures looked connected
The outages arrived on the fourth straight day of major disruptions, but the available explanations pointed to separate technical problems. The FAA called its incident “routine maintenance” involving the replacement of copper communications lines with fiber-optic cable. FAA head Bryan Bedford said, “We see zero evidence of anything nefarious.”
Microsoft blamed an old software bug, malfunctioning internal data connections, and server capacity. JPMorgan Chase attributed its outage to a technical issue, while Apple blamed an internal update. The NSA said problems arose “during scheduled routine maintenance.” US Secretary of State Marco Rubio said there was “zero evidence or any reason to believe” the isolated incidents involved hackers, criminals, another nation state, or any other malign actor.
Robert N. Charette called the timing unusual, not the outages themselves. “Over the last 20 years of reporting for Spectrum, I have documented hundreds of outages over the years,” he said. “To my mind, none of these are uncommon. What is uncommon is that they’re clustered together.”
Researchers described a broader infrastructure problem. Daniel Dresner of the University of Manchester said, “Things are failing all the time,” and warned, “This stuff is really fragile.” He argued that infrastructure lacks investment and said, “We wanted agility and we ended up with fragility.”
Tim Stevens of King’s College London said the technology sector has built profitable, desirable tools without maintaining the infrastructure beneath them. He said AI has helped uncover many software bugs, which have required many patches, while the economics of cybersecurity make thorough testing difficult. George Loukas of the University of Greenwich said people depend on a small number of digital tools and that digital environments tend to be an oligopoly.
Charette also warned about the race between deploying AI and maintaining older systems. He said organizations and governments may be taking money for AI from legacy systems’ operational maintenance budgets. Adding AI to systems that operators do not understand well, he said, could create more failures because risk exists at the interfaces between systems.
Alon Hillel-Tuch of New York University offered a simpler explanation: “Outages like these happen all the time.” Tim Stevens said the current cluster looked like coincidence rather than one coordinated event. The lesson from the network research is just as practical: reliable systems depend on their structure, their routing rules, their maintenance, and the links between every layer.
As JournalBot wrote on Ars OpenForum, “When everything goes down, don’t assume it’s a cyberattack.”
Based on
- Network Researchers Nail Down Theory of ‘Perfect Resilience’ — quantamagazine.org
- What the IT Outage Cluster Really Reveals – IEEE Spectrum — spectrum.ieee.org
- Global tech outages raise questions about vulnerabilities in digital infrastructure – CTVNews — ctvnews.ca
- The real reasons we keep seeing big tech outages – it isn’t just hackers | New Scientist — newscientist.com
- How to tell a cyberattack from a plain old outage | Ars OpenForum — arstechnica.com




